In a mobile communication system, speech signals are requested to be compressed into low bit rates for transmission so as to efficiently utilize radio wave resources and so on. On the other hand, improvement of speech call audio quality and achievement of high quality realistic speech call service are desired. To achieve them, not only a mono signal but also multi-channel audio signals, in particular stereo audio signals, are desirably encoded with high quality.
A method using correlation between channels is effective to encode stereo audio signal (two-channel audio signals) or multi-channel audio signals with low bit rates. A method for backward adaptive prediction of a signal in a channel from a signal in another channel using an adaptive filter is known as a method using correlation between channels (see non-Patent Literature 1 and Patent Literature 1).
In this method, when a signal reaches a left microphone and a right microphone from a sound source, acoustic characteristics between a sound source-a left microphone and between the sound source-a right microphone is estimated using an adaptive filter. A FIR (Finite Impulse Response) filter is used as the adaptive filter.
An estimation method using the adaptive filter will be hereinafter explained using an example where acoustic characteristic of a stereo audio signal are estimated.
In FIG. 1, HL(z) represents acoustic characteristic between a sound source and a left microphone, and HR(z) represents acoustic characteristic between the sound source and a right microphone. If the right signal is estimated from the left signal using the adaptive filter, a transfer function G(z) of the adaptive filter is configured to satisfy the relationship of equation 1 with regard to HL(z) and HR(z).
                              (                      Equation            ⁢                                                  ⁢            1                    )                ⁢                                                                                                G          ⁡                      (            z            )                          =                                            H              R                        ⁡                          (              z              )                                                          H              L                        ⁡                          (              z              )                                                          [        1        ]            
Using the adaptive filter having the transfer function G(z) satisfying equation 1, the right signal is estimated from the left signal, and the estimated error is quantized. In this manner, using the adaptive filter, the correlation between the left signal and the right signal is removed, whereby efficient encoding can be achieved.
The transfer function G(z) of the adaptive filter is expressed as equation 2.
                              (                      Equation            ⁢                                                  ⁢            2                    )                ⁢                                                                                                G          ⁡                      (            z            )                          =                              ∑                          n              =              0                                      N              -              1                                ⁢                                                    g                k                            ⁡                              (                n                )                                      ·                          z                              -                n                                                                        [        2        ]            
In equation 2, gk(n) denotes the n-th (filter coefficient order n) filter coefficient of the adaptive filter at time k, z denotes a Z-transformation variable, and N denotes a filter order of the adaptive filter (the maximum value of filter coefficient order n).
The adaptive filter estimates acoustic characteristic while successively updating the filter coefficient in units of sample processings. When learning identification method (NLMS (normalized least-mean-square)) algorithm is used to update the filter coefficient of the adaptive filter, filter coefficient gk(n) of the adaptive filter is updated according to equation 3.
                              (                      Equation            ⁢                                                  ⁢            3                    )                ⁢                                                                                                            g                          k              +              1                                ⁡                      (            n            )                          =                                            g              k                        ⁡                          (              n              )                                +                                                    α                                                                            ∑                                              i                        =                        0                                                                    N                        -                        1                                                              ⁢                                                                                            x                          k                                                ⁡                                                  (                          i                          )                                                                    2                                                        +                  β                                            ·                              e                ⁡                                  (                  k                  )                                            ·                                                x                  k                                ⁡                                  (                  n                  )                                                      ⁢                                                  ⁢                          (                              for                ⁢                                                                  ⁢                all                ⁢                                                                  ⁢                n                            )                                                          [        3        ]            
As described above, gk(n) denotes the n-th (filter coefficient order n) filter coefficient of the adaptive filter at time k, and N denotes the filter order of the adaptive filter (the maximum value of filter coefficient order n). On the other hand, e(k) denotes an error signal at time k, and xk(n) denotes an input signal at time k multiplied by the n-th (filter coefficient order n) filter coefficient of the adaptive filter. α denotes a parameter for controlling update speed of the adaptive filter, and β denotes a parameter for preventing a denominator of equation 3 from being zero. β is a positive value.
At this occasion, the filter order N of the adaptive filter needs to be determined according to acoustic characteristic between the sound source and the microphone. For example, it is necessary to represent acoustic characteristic for a time length of about 100 ms in order to ensure sufficient performance. In this case, the filter coefficient of the adaptive filter needs to have a filter order N for the time length of 100 ms. Accordingly, when the sampling frequency of the input signal is 32 kHz, the filter order N of the adaptive filter required to obtain the acoustic characteristic for the time length of 100 ms is 3200.